+ return thread_block_reason(continuation, parameter, AST_NONE);
+}
+
+/*
+ * thread_run:
+ *
+ * Switch directly from the current thread to the
+ * new thread, handing off our quantum if appropriate.
+ *
+ * New thread must be runnable, and not on a run queue.
+ *
+ * Called at splsched.
+ */
+int
+thread_run(
+ thread_t self,
+ thread_continue_t continuation,
+ void *parameter,
+ thread_t new_thread)
+{
+ ast_t reason = AST_HANDOFF;
+
+ self->continuation = continuation;
+ self->parameter = parameter;
+
+ while (!thread_invoke(self, new_thread, reason)) {
+ /* the handoff failed, so we have to fall back to the normal block path */
+ processor_t processor = current_processor();
+
+ reason = AST_NONE;
+
+ thread_lock(self);
+ new_thread = thread_select(self, processor, &reason);
+ thread_unlock(self);
+ }
+
+ return (self->wait_result);
+}
+
+/*
+ * thread_continue:
+ *
+ * Called at splsched when a thread first receives
+ * a new stack after a continuation.
+ */
+void
+thread_continue(
+ thread_t thread)
+{
+ thread_t self = current_thread();
+ thread_continue_t continuation;
+ void *parameter;
+
+ DTRACE_SCHED(on__cpu);
+
+ continuation = self->continuation;
+ parameter = self->parameter;
+
+#if KPERF
+ kperf_on_cpu(self, continuation, NULL);
+#endif
+
+ thread_dispatch(thread, self);
+
+ self->continuation = self->parameter = NULL;
+
+#if INTERRUPT_MASKED_DEBUG
+ /* Reset interrupt-masked spin debugging timeout */
+ ml_spin_debug_clear(self);
+#endif
+
+ TLOG(1, "thread_continue: calling call_continuation\n");
+
+ boolean_t enable_interrupts = thread != THREAD_NULL;
+ call_continuation(continuation, parameter, self->wait_result, enable_interrupts);
+ /*NOTREACHED*/
+}
+
+void
+thread_quantum_init(thread_t thread)
+{
+ if (thread->sched_mode == TH_MODE_REALTIME) {
+ thread->quantum_remaining = thread->realtime.computation;
+ } else {
+ thread->quantum_remaining = SCHED(initial_quantum_size)(thread);
+ }
+}
+
+uint32_t
+sched_timeshare_initial_quantum_size(thread_t thread)
+{
+ if ((thread != THREAD_NULL) && thread->th_sched_bucket == TH_BUCKET_SHARE_BG)
+ return bg_quantum;
+ else
+ return std_quantum;
+}
+
+/*
+ * run_queue_init:
+ *
+ * Initialize a run queue before first use.
+ */
+void
+run_queue_init(
+ run_queue_t rq)
+{
+ rq->highq = NOPRI;
+ for (u_int i = 0; i < BITMAP_LEN(NRQS); i++)
+ rq->bitmap[i] = 0;
+ rq->urgency = rq->count = 0;
+ for (int i = 0; i < NRQS; i++)
+ queue_init(&rq->queues[i]);
+}
+
+/*
+ * run_queue_dequeue:
+ *
+ * Perform a dequeue operation on a run queue,
+ * and return the resulting thread.
+ *
+ * The run queue must be locked (see thread_run_queue_remove()
+ * for more info), and not empty.
+ */
+thread_t
+run_queue_dequeue(
+ run_queue_t rq,
+ integer_t options)
+{
+ thread_t thread;
+ queue_t queue = &rq->queues[rq->highq];
+
+ if (options & SCHED_HEADQ) {
+ thread = qe_dequeue_head(queue, struct thread, runq_links);
+ } else {
+ thread = qe_dequeue_tail(queue, struct thread, runq_links);
+ }
+
+ assert(thread != THREAD_NULL);
+ assert_thread_magic(thread);
+
+ thread->runq = PROCESSOR_NULL;
+ SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
+ rq->count--;
+ if (SCHED(priority_is_urgent)(rq->highq)) {
+ rq->urgency--; assert(rq->urgency >= 0);
+ }
+ if (queue_empty(queue)) {
+ bitmap_clear(rq->bitmap, rq->highq);
+ rq->highq = bitmap_first(rq->bitmap, NRQS);
+ }
+
+ return thread;
+}
+
+/*
+ * run_queue_enqueue:
+ *
+ * Perform a enqueue operation on a run queue.
+ *
+ * The run queue must be locked (see thread_run_queue_remove()
+ * for more info).
+ */
+boolean_t
+run_queue_enqueue(
+ run_queue_t rq,
+ thread_t thread,
+ integer_t options)
+{
+ queue_t queue = &rq->queues[thread->sched_pri];
+ boolean_t result = FALSE;
+
+ assert_thread_magic(thread);
+
+ if (queue_empty(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+
+ rq_bitmap_set(rq->bitmap, thread->sched_pri);
+ if (thread->sched_pri > rq->highq) {
+ rq->highq = thread->sched_pri;
+ result = TRUE;
+ }
+ } else {
+ if (options & SCHED_TAILQ)
+ enqueue_tail(queue, &thread->runq_links);
+ else
+ enqueue_head(queue, &thread->runq_links);
+ }
+ if (SCHED(priority_is_urgent)(thread->sched_pri))
+ rq->urgency++;
+ SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
+ rq->count++;
+
+ return (result);
+}
+
+/*
+ * run_queue_remove:
+ *
+ * Remove a specific thread from a runqueue.
+ *
+ * The run queue must be locked.
+ */
+void
+run_queue_remove(
+ run_queue_t rq,
+ thread_t thread)
+{
+ assert(thread->runq != PROCESSOR_NULL);
+ assert_thread_magic(thread);
+
+ remqueue(&thread->runq_links);
+ SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
+ rq->count--;
+ if (SCHED(priority_is_urgent)(thread->sched_pri)) {
+ rq->urgency--; assert(rq->urgency >= 0);
+ }
+
+ if (queue_empty(&rq->queues[thread->sched_pri])) {
+ /* update run queue status */
+ bitmap_clear(rq->bitmap, thread->sched_pri);
+ rq->highq = bitmap_first(rq->bitmap, NRQS);
+ }
+
+ thread->runq = PROCESSOR_NULL;
+}
+
+/* Assumes RT lock is not held, and acquires splsched/rt_lock itself */
+void
+sched_rtglobal_runq_scan(sched_update_scan_context_t scan_context)
+{
+ spl_t s;
+ thread_t thread;
+
+ processor_set_t pset = &pset0;
+
+ s = splsched();
+ rt_lock_lock(pset);
+
+ qe_foreach_element_safe(thread, &pset->rt_runq.queue, runq_links) {
+ if (thread->last_made_runnable_time < scan_context->earliest_rt_make_runnable_time) {
+ scan_context->earliest_rt_make_runnable_time = thread->last_made_runnable_time;
+ }
+ }
+
+ rt_lock_unlock(pset);
+ splx(s);
+}
+
+int64_t
+sched_rtglobal_runq_count_sum(void)
+{
+ return pset0.rt_runq.runq_stats.count_sum;
+}
+
+/*
+ * realtime_queue_insert:
+ *
+ * Enqueue a thread for realtime execution.
+ */
+static boolean_t
+realtime_queue_insert(processor_t processor, processor_set_t pset, thread_t thread)
+{
+ queue_t queue = &SCHED(rt_runq)(pset)->queue;
+ uint64_t deadline = thread->realtime.deadline;
+ boolean_t preempt = FALSE;
+
+ rt_lock_lock(pset);
+
+ if (queue_empty(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+ preempt = TRUE;
+ } else {
+ /* Insert into rt_runq in thread deadline order */
+ queue_entry_t iter;
+ qe_foreach(iter, queue) {
+ thread_t iter_thread = qe_element(iter, struct thread, runq_links);
+ assert_thread_magic(iter_thread);
+
+ if (deadline < iter_thread->realtime.deadline) {
+ if (iter == queue_first(queue))
+ preempt = TRUE;
+ insque(&thread->runq_links, queue_prev(iter));
+ break;
+ } else if (iter == queue_last(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+ break;
+ }
+ }
+ }
+
+ thread->runq = processor;
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_incr(pset);
+
+ rt_lock_unlock(pset);
+
+ return (preempt);
+}
+
+/*
+ * realtime_setrun:
+ *
+ * Dispatch a thread for realtime execution.
+ *
+ * Thread must be locked. Associated pset must
+ * be locked, and is returned unlocked.
+ */
+static void
+realtime_setrun(
+ processor_t processor,
+ thread_t thread)
+{
+ processor_set_t pset = processor->processor_set;
+ pset_assert_locked(pset);
+ ast_t preempt;
+
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ thread->chosen_processor = processor;
+
+ /* <rdar://problem/15102234> */
+ assert(thread->bound_processor == PROCESSOR_NULL);
+
+ /*
+ * Dispatch directly onto idle processor.
+ */
+ if ( (thread->bound_processor == processor)
+ && processor->state == PROCESSOR_IDLE) {
+
+ processor->next_thread = thread;
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+
+ ipi_type = sched_ipi_action(processor, thread, true, SCHED_IPI_EVENT_BOUND_THR);
+ pset_unlock(pset);
+ sched_ipi_perform(processor, ipi_type);
+ return;
+ }
+
+ if (processor->current_pri < BASEPRI_RTQUEUES)
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else if (thread->realtime.deadline < processor->deadline)
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else
+ preempt = AST_NONE;
+
+ realtime_queue_insert(processor, pset, thread);
+
+ ipi_type = SCHED_IPI_NONE;
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ processor->next_thread = THREAD_NULL;
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+ if (processor == current_processor()) {
+ ast_on(preempt);
+ } else {
+ ipi_type = sched_ipi_action(processor, thread, true, SCHED_IPI_EVENT_PREEMPT);
+ }
+ } else if (processor->state == PROCESSOR_DISPATCHING) {
+ if ((processor->next_thread == THREAD_NULL) && ((processor->current_pri < thread->sched_pri) || (processor->deadline > thread->realtime.deadline))) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ }
+ } else {
+ if (processor == current_processor()) {
+ ast_on(preempt);
+ } else {
+ ipi_type = sched_ipi_action(processor, thread, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ } else {
+ /* Selected processor was too busy, just keep thread enqueued and let other processors drain it naturally. */
+ }
+
+ pset_unlock(pset);
+ sched_ipi_perform(processor, ipi_type);
+}
+
+
+sched_ipi_type_t sched_ipi_deferred_policy(processor_set_t pset, processor_t dst,
+ __unused sched_ipi_event_t event)
+{
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ if (!bit_test(pset->pending_deferred_AST_cpu_mask, dst->cpu_id)) {
+ return SCHED_IPI_DEFERRED;
+ }
+#else /* CONFIG_SCHED_DEFERRED_AST */
+ panic("Request for deferred IPI on an unsupported platform; pset: %p CPU: %d", pset, dst->cpu_id);
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+ return SCHED_IPI_NONE;
+}
+
+sched_ipi_type_t sched_ipi_action(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event)
+{
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ assert(dst != NULL);
+
+ processor_set_t pset = dst->processor_set;
+ if (current_processor() == dst) {
+ return SCHED_IPI_NONE;
+ }
+
+ if (bit_test(pset->pending_AST_cpu_mask, dst->cpu_id)) {
+ return SCHED_IPI_NONE;
+ }
+
+ ipi_type = SCHED(ipi_policy)(dst, thread, dst_idle, event);
+ switch(ipi_type) {
+ case SCHED_IPI_NONE:
+ return SCHED_IPI_NONE;
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ case SCHED_IPI_DEFERRED:
+ bit_set(pset->pending_deferred_AST_cpu_mask, dst->cpu_id);
+ break;
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+ default:
+ bit_set(pset->pending_AST_cpu_mask, dst->cpu_id);
+ break;
+ }
+ return ipi_type;
+}
+
+sched_ipi_type_t sched_ipi_policy(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event)
+{
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ boolean_t deferred_ipi_supported = false;
+ processor_set_t pset = dst->processor_set;
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ deferred_ipi_supported = true;
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+
+ switch(event) {
+ case SCHED_IPI_EVENT_SPILL:
+ case SCHED_IPI_EVENT_SMT_REBAL:
+ case SCHED_IPI_EVENT_REBALANCE:
+ case SCHED_IPI_EVENT_BOUND_THR:
+ /*
+ * The spill, SMT rebalance, rebalance and the bound thread
+ * scenarios use immediate IPIs always.
+ */
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ case SCHED_IPI_EVENT_PREEMPT:
+ /* In the preemption case, use immediate IPIs for RT threads */
+ if (thread && (thread->sched_pri >= BASEPRI_RTQUEUES)) {
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ }
+
+ /*
+ * For Non-RT threads preemption,
+ * If the core is active, use immediate IPIs.
+ * If the core is idle, use deferred IPIs if supported; otherwise immediate IPI.
+ */
+ if (deferred_ipi_supported && dst_idle) {
+ return sched_ipi_deferred_policy(pset, dst, event);
+ }
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ default:
+ panic("Unrecognized scheduler IPI event type %d", event);
+ }
+ assert(ipi_type != SCHED_IPI_NONE);
+ return ipi_type;
+}
+
+void sched_ipi_perform(processor_t dst, sched_ipi_type_t ipi)
+{
+ switch (ipi) {
+ case SCHED_IPI_NONE:
+ break;
+ case SCHED_IPI_IDLE:
+ machine_signal_idle(dst);
+ break;
+ case SCHED_IPI_IMMEDIATE:
+ cause_ast_check(dst);
+ break;
+ case SCHED_IPI_DEFERRED:
+ machine_signal_idle_deferred(dst);
+ break;
+ default:
+ panic("Unrecognized scheduler IPI type: %d", ipi);
+ }
+}
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+
+boolean_t
+priority_is_urgent(int priority)
+{
+ return bitmap_test(sched_preempt_pri, priority) ? TRUE : FALSE;
+}
+
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+
+/*
+ * processor_setrun:
+ *
+ * Dispatch a thread for execution on a
+ * processor.
+ *
+ * Thread must be locked. Associated pset must
+ * be locked, and is returned unlocked.
+ */
+static void
+processor_setrun(
+ processor_t processor,
+ thread_t thread,
+ integer_t options)
+{
+ processor_set_t pset = processor->processor_set;
+ pset_assert_locked(pset);
+ ast_t preempt;
+ enum { eExitIdle, eInterruptRunning, eDoNothing } ipi_action = eDoNothing;
+
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ thread->chosen_processor = processor;
+
+ /*
+ * Dispatch directly onto idle processor.
+ */
+ if ( (SCHED(direct_dispatch_to_idle_processors) ||
+ thread->bound_processor == processor)
+ && processor->state == PROCESSOR_IDLE) {
+
+ processor->next_thread = thread;
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+
+ ipi_type = sched_ipi_action(processor, thread, true, SCHED_IPI_EVENT_BOUND_THR);
+ pset_unlock(pset);
+ sched_ipi_perform(processor, ipi_type);
+ return;
+ }
+
+ /*
+ * Set preemption mode.
+ */
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ /* TODO: Do we need to care about urgency (see rdar://problem/20136239)? */
+#endif
+ if (SCHED(priority_is_urgent)(thread->sched_pri) && thread->sched_pri > processor->current_pri)
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else if(processor->active_thread && thread_eager_preemption(processor->active_thread))
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else if ((thread->sched_mode == TH_MODE_TIMESHARE) && (thread->sched_pri < thread->base_pri)) {
+ if(SCHED(priority_is_urgent)(thread->base_pri) && thread->sched_pri > processor->current_pri) {
+ preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+ } else {
+ preempt = AST_NONE;
+ }
+ } else
+ preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+
+ if ((options & (SCHED_PREEMPT|SCHED_REBALANCE)) == (SCHED_PREEMPT|SCHED_REBALANCE)) {
+ /*
+ * Having gone to the trouble of forcing this thread off a less preferred core,
+ * we should force the preferable core to reschedule immediately to give this
+ * thread a chance to run instead of just sitting on the run queue where
+ * it may just be stolen back by the idle core we just forced it off.
+ */
+ preempt |= AST_PREEMPT;
+ }
+
+ SCHED(processor_enqueue)(processor, thread, options);
+ sched_update_pset_load_average(pset);
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ processor->next_thread = THREAD_NULL;
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+ ipi_action = eExitIdle;
+ } else if ( processor->state == PROCESSOR_DISPATCHING) {
+ if ((processor->next_thread == THREAD_NULL) && (processor->current_pri < thread->sched_pri)) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ }
+ } else if ( (processor->state == PROCESSOR_RUNNING ||
+ processor->state == PROCESSOR_SHUTDOWN) &&
+ (thread->sched_pri >= processor->current_pri)) {
+ ipi_action = eInterruptRunning;
+ }
+ } else {
+ /*
+ * New thread is not important enough to preempt what is running, but
+ * special processor states may need special handling
+ */
+ if (processor->state == PROCESSOR_SHUTDOWN &&
+ thread->sched_pri >= processor->current_pri ) {
+ ipi_action = eInterruptRunning;
+ } else if (processor->state == PROCESSOR_IDLE) {
+
+ processor->next_thread = THREAD_NULL;
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+
+ ipi_action = eExitIdle;
+ }
+ }
+
+ if (ipi_action != eDoNothing) {
+ if (processor == current_processor()) {
+ if (csw_check_locked(processor, pset, AST_NONE) != AST_NONE)
+ ast_on(preempt);
+ } else {
+ sched_ipi_event_t event = (options & SCHED_REBALANCE) ? SCHED_IPI_EVENT_REBALANCE : SCHED_IPI_EVENT_PREEMPT;
+ ipi_type = sched_ipi_action(processor, thread, (ipi_action == eExitIdle), event);
+ }
+ }
+ pset_unlock(pset);
+ sched_ipi_perform(processor, ipi_type);
+}
+
+/*
+ * choose_next_pset:
+ *
+ * Return the next sibling pset containing
+ * available processors.
+ *
+ * Returns the original pset if none other is
+ * suitable.
+ */
+static processor_set_t
+choose_next_pset(
+ processor_set_t pset)
+{
+ processor_set_t nset = pset;
+
+ do {
+ nset = next_pset(nset);
+ } while (nset->online_processor_count < 1 && nset != pset);
+
+ return (nset);
+}
+
+/*
+ * choose_processor:
+ *
+ * Choose a processor for the thread, beginning at
+ * the pset. Accepts an optional processor hint in
+ * the pset.
+ *
+ * Returns a processor, possibly from a different pset.
+ *
+ * The thread must be locked. The pset must be locked,
+ * and the resulting pset is locked on return.
+ */
+processor_t
+choose_processor(
+ processor_set_t starting_pset,
+ processor_t processor,
+ thread_t thread)
+{
+ processor_set_t pset = starting_pset;
+ processor_set_t nset;
+
+ assert(thread->sched_pri <= BASEPRI_RTQUEUES);
+
+ /*
+ * Prefer the hinted processor, when appropriate.
+ */
+
+ /* Fold last processor hint from secondary processor to its primary */
+ if (processor != PROCESSOR_NULL) {
+ processor = processor->processor_primary;
+ }
+
+ /*
+ * Only consult platform layer if pset is active, which
+ * it may not be in some cases when a multi-set system
+ * is going to sleep.
+ */
+ if (pset->online_processor_count) {
+ if ((processor == PROCESSOR_NULL) || (processor->processor_set == pset && processor->state == PROCESSOR_IDLE)) {
+ processor_t mc_processor = machine_choose_processor(pset, processor);
+ if (mc_processor != PROCESSOR_NULL)
+ processor = mc_processor->processor_primary;
+ }
+ }
+
+ /*
+ * At this point, we may have a processor hint, and we may have
+ * an initial starting pset. If the hint is not in the pset, or
+ * if the hint is for a processor in an invalid state, discard
+ * the hint.
+ */
+ if (processor != PROCESSOR_NULL) {
+ if (processor->processor_set != pset) {
+ processor = PROCESSOR_NULL;
+ } else if (!processor->is_recommended) {
+ processor = PROCESSOR_NULL;
+ } else {
+ switch (processor->state) {
+ case PROCESSOR_START:
+ case PROCESSOR_SHUTDOWN:
+ case PROCESSOR_OFF_LINE:
+ /*
+ * Hint is for a processor that cannot support running new threads.
+ */
+ processor = PROCESSOR_NULL;
+ break;
+ case PROCESSOR_IDLE:
+ /*
+ * Hint is for an idle processor. Assume it is no worse than any other
+ * idle processor. The platform layer had an opportunity to provide
+ * the "least cost idle" processor above.
+ */
+ return (processor);
+ case PROCESSOR_RUNNING:
+ case PROCESSOR_DISPATCHING:
+ /*
+ * Hint is for an active CPU. This fast-path allows
+ * realtime threads to preempt non-realtime threads
+ * to regain their previous executing processor.
+ */
+ if ((thread->sched_pri >= BASEPRI_RTQUEUES) &&
+ (processor->current_pri < BASEPRI_RTQUEUES))
+ return (processor);
+
+ /* Otherwise, use hint as part of search below */
+ break;
+ default:
+ processor = PROCESSOR_NULL;
+ break;
+ }
+ }
+ }
+
+ /*
+ * Iterate through the processor sets to locate
+ * an appropriate processor. Seed results with
+ * a last-processor hint, if available, so that
+ * a search must find something strictly better
+ * to replace it.
+ *
+ * A primary/secondary pair of SMT processors are
+ * "unpaired" if the primary is busy but its
+ * corresponding secondary is idle (so the physical
+ * core has full use of its resources).
+ */
+
+ integer_t lowest_priority = MAXPRI + 1;
+ integer_t lowest_secondary_priority = MAXPRI + 1;
+ integer_t lowest_unpaired_primary_priority = MAXPRI + 1;
+ integer_t lowest_count = INT_MAX;
+ uint64_t furthest_deadline = 1;
+ processor_t lp_processor = PROCESSOR_NULL;
+ processor_t lp_unpaired_primary_processor = PROCESSOR_NULL;
+ processor_t lp_unpaired_secondary_processor = PROCESSOR_NULL;
+ processor_t lp_paired_secondary_processor = PROCESSOR_NULL;
+ processor_t lc_processor = PROCESSOR_NULL;
+ processor_t fd_processor = PROCESSOR_NULL;
+
+ if (processor != PROCESSOR_NULL) {
+ /* All other states should be enumerated above. */
+ assert(processor->state == PROCESSOR_RUNNING || processor->state == PROCESSOR_DISPATCHING);
+
+ lowest_priority = processor->current_pri;
+ lp_processor = processor;
+
+ if (processor->current_pri >= BASEPRI_RTQUEUES) {
+ furthest_deadline = processor->deadline;
+ fd_processor = processor;
+ }
+
+ lowest_count = SCHED(processor_runq_count)(processor);
+ lc_processor = processor;
+ }
+
+ do {
+ /*
+ * Choose an idle processor, in pset traversal order
+ */
+
+ uint64_t idle_primary_map = (pset->cpu_state_map[PROCESSOR_IDLE] &
+ pset->primary_map &
+ pset->recommended_bitmask &
+ ~pset->pending_AST_cpu_mask);
+
+ int cpuid = lsb_first(idle_primary_map);
+ if (cpuid >= 0) {
+ processor = processor_array[cpuid];
+ return processor;
+ }
+
+ /*
+ * Otherwise, enumerate active and idle processors to find primary candidates
+ * with lower priority/etc.
+ */
+
+ uint64_t active_map = ((pset->cpu_state_map[PROCESSOR_RUNNING] | pset->cpu_state_map[PROCESSOR_DISPATCHING]) &
+ pset->recommended_bitmask &
+ ~pset->pending_AST_cpu_mask);
+ active_map = bit_ror64(active_map, (pset->last_chosen + 1));
+ for (int rotid = lsb_first(active_map); rotid >= 0; rotid = lsb_next(active_map, rotid)) {
+ cpuid = ((rotid + pset->last_chosen + 1) & 63);
+ processor = processor_array[cpuid];
+
+ integer_t cpri = processor->current_pri;
+ if (processor->processor_primary != processor) {
+ if (cpri < lowest_secondary_priority) {
+ lowest_secondary_priority = cpri;
+ lp_paired_secondary_processor = processor;
+ }
+ } else {
+ if (cpri < lowest_priority) {
+ lowest_priority = cpri;
+ lp_processor = processor;
+ }
+ }
+
+ if ((cpri >= BASEPRI_RTQUEUES) && (processor->deadline > furthest_deadline)) {
+ furthest_deadline = processor->deadline;
+ fd_processor = processor;
+ }
+
+ integer_t ccount = SCHED(processor_runq_count)(processor);
+ if (ccount < lowest_count) {
+ lowest_count = ccount;
+ lc_processor = processor;
+ }
+ }
+
+ /*
+ * For SMT configs, these idle secondary processors must have active primary. Otherwise
+ * the idle primary would have short-circuited the loop above
+ */
+ uint64_t idle_secondary_map = (pset->cpu_state_map[PROCESSOR_IDLE] &
+ ~pset->primary_map &
+ pset->recommended_bitmask &
+ ~pset->pending_AST_cpu_mask);
+
+ for (cpuid = lsb_first(idle_secondary_map); cpuid >= 0; cpuid = lsb_next(idle_secondary_map, cpuid)) {
+ processor = processor_array[cpuid];
+
+ processor_t cprimary = processor->processor_primary;
+
+ if (!cprimary->is_recommended) {
+ continue;
+ }
+ if (bit_test(pset->pending_AST_cpu_mask, cprimary->cpu_id)) {
+ continue;
+ }
+
+ /* If the primary processor is offline or starting up, it's not a candidate for this path */
+ if (cprimary->state == PROCESSOR_RUNNING || cprimary->state == PROCESSOR_DISPATCHING) {
+ integer_t primary_pri = cprimary->current_pri;
+
+ if (primary_pri < lowest_unpaired_primary_priority) {
+ lowest_unpaired_primary_priority = primary_pri;
+ lp_unpaired_primary_processor = cprimary;
+ lp_unpaired_secondary_processor = processor;
+ }
+ }
+ }
+
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+
+ /*
+ * For realtime threads, the most important aspect is
+ * scheduling latency, so we attempt to assign threads
+ * to good preemption candidates (assuming an idle primary
+ * processor was not available above).
+ */
+
+ if (thread->sched_pri > lowest_unpaired_primary_priority) {
+ pset->last_chosen = lp_unpaired_primary_processor->cpu_id;
+ return lp_unpaired_primary_processor;
+ }
+ if (thread->sched_pri > lowest_priority) {
+ pset->last_chosen = lp_processor->cpu_id;
+ return lp_processor;
+ }
+ if (sched_allow_rt_smt && (thread->sched_pri > lowest_secondary_priority)) {
+ pset->last_chosen = lp_paired_secondary_processor->cpu_id;
+ return lp_paired_secondary_processor;
+ }
+ if (thread->realtime.deadline < furthest_deadline)
+ return fd_processor;
+
+ /*
+ * If all primary and secondary CPUs are busy with realtime
+ * threads with deadlines earlier than us, move on to next
+ * pset.
+ */
+ }
+ else {
+
+ if (thread->sched_pri > lowest_unpaired_primary_priority) {
+ pset->last_chosen = lp_unpaired_primary_processor->cpu_id;
+ return lp_unpaired_primary_processor;
+ }
+ if (thread->sched_pri > lowest_priority) {
+ pset->last_chosen = lp_processor->cpu_id;
+ return lp_processor;
+ }
+
+ /*
+ * If all primary processor in this pset are running a higher
+ * priority thread, move on to next pset. Only when we have
+ * exhausted this search do we fall back to other heuristics.
+ */
+ }
+
+ /*
+ * Move onto the next processor set.
+ */
+ nset = next_pset(pset);
+
+ if (nset != starting_pset) {
+ pset_unlock(pset);
+
+ pset = nset;
+ pset_lock(pset);
+ }
+ } while (nset != starting_pset);
+
+ /*
+ * Make sure that we pick a running processor,
+ * and that the correct processor set is locked.
+ * Since we may have unlock the candidate processor's
+ * pset, it may have changed state.
+ *
+ * All primary processors are running a higher priority
+ * thread, so the only options left are enqueuing on
+ * the secondary processor that would perturb the least priority
+ * primary, or the least busy primary.
+ */
+ do {
+
+ /* lowest_priority is evaluated in the main loops above */
+ if (lp_unpaired_secondary_processor != PROCESSOR_NULL) {
+ processor = lp_unpaired_secondary_processor;
+ lp_unpaired_secondary_processor = PROCESSOR_NULL;
+ } else if (lp_paired_secondary_processor != PROCESSOR_NULL) {
+ processor = lp_paired_secondary_processor;
+ lp_paired_secondary_processor = PROCESSOR_NULL;
+ } else if (lc_processor != PROCESSOR_NULL) {
+ processor = lc_processor;
+ lc_processor = PROCESSOR_NULL;
+ } else {
+ /*
+ * All processors are executing higher
+ * priority threads, and the lowest_count
+ * candidate was not usable
+ */
+ processor = master_processor;
+ }
+
+ /*
+ * Check that the correct processor set is
+ * returned locked.
+ */
+ if (pset != processor->processor_set) {
+ pset_unlock(pset);
+ pset = processor->processor_set;
+ pset_lock(pset);
+ }
+
+ /*
+ * We must verify that the chosen processor is still available.
+ * master_processor is an exception, since we may need to preempt
+ * a running thread on it during processor shutdown (for sleep),
+ * and that thread needs to be enqueued on its runqueue to run
+ * when the processor is restarted.
+ */
+ if (processor != master_processor && (processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE))
+ processor = PROCESSOR_NULL;
+
+ } while (processor == PROCESSOR_NULL);
+
+ pset->last_chosen = processor->cpu_id;
+ return processor;
+}
+
+/*
+ * thread_setrun:
+ *
+ * Dispatch thread for execution, onto an idle
+ * processor or run queue, and signal a preemption
+ * as appropriate.
+ *
+ * Thread must be locked.
+ */
+void
+thread_setrun(
+ thread_t thread,
+ integer_t options)
+{
+ processor_t processor;
+ processor_set_t pset;
+
+ assert((thread->state & (TH_RUN|TH_WAIT|TH_UNINT|TH_TERMINATE|TH_TERMINATE2)) == TH_RUN);
+ assert(thread->runq == PROCESSOR_NULL);
+
+ /*
+ * Update priority if needed.
+ */
+ if (SCHED(can_update_priority)(thread))
+ SCHED(update_priority)(thread);
+
+ thread->sfi_class = sfi_thread_classify(thread);
+
+ assert(thread->runq == PROCESSOR_NULL);
+
+#if __SMP__
+ if (thread->bound_processor == PROCESSOR_NULL) {
+ /*
+ * Unbound case.
+ */
+ if (thread->affinity_set != AFFINITY_SET_NULL) {
+ /*
+ * Use affinity set policy hint.
+ */
+ pset = thread->affinity_set->aset_pset;
+ pset_lock(pset);
+
+ processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread);
+ pset = processor->processor_set;
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR)|DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-1, processor->cpu_id, processor->state, 0);
+ } else if (thread->last_processor != PROCESSOR_NULL) {
+ /*
+ * Simple (last processor) affinity case.
+ */
+ processor = thread->last_processor;
+ pset = processor->processor_set;
+ pset_lock(pset);
+ processor = SCHED(choose_processor)(pset, processor, thread);
+ pset = processor->processor_set;
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR)|DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->last_processor->cpu_id, processor->cpu_id, processor->state, 0);
+ } else {
+ /*
+ * No Affinity case:
+ *
+ * Utilitize a per task hint to spread threads
+ * among the available processor sets.
+ */
+ task_t task = thread->task;
+
+ pset = task->pset_hint;
+ if (pset == PROCESSOR_SET_NULL)
+ pset = current_processor()->processor_set;
+
+ pset = choose_next_pset(pset);
+ pset_lock(pset);
+
+ processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread);
+ pset = processor->processor_set;
+ task->pset_hint = pset;
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR)|DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-1, processor->cpu_id, processor->state, 0);
+ }
+ } else {
+ /*
+ * Bound case:
+ *
+ * Unconditionally dispatch on the processor.
+ */
+ processor = thread->bound_processor;
+ pset = processor->processor_set;
+ pset_lock(pset);
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR)|DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-2, processor->cpu_id, processor->state, 0);
+ }
+#else /* !__SMP__ */
+ /* Only one processor to choose */
+ assert(thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == master_processor);
+ processor = master_processor;
+ pset = processor->processor_set;
+ pset_lock(pset);
+#endif /* !__SMP__ */
+
+ /*
+ * Dispatch the thread on the chosen processor.
+ * TODO: This should be based on sched_mode, not sched_pri
+ */
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ realtime_setrun(processor, thread);
+ } else {
+ processor_setrun(processor, thread, options);
+ }
+ /* pset is now unlocked */
+ if (thread->bound_processor == PROCESSOR_NULL) {
+ SCHED(check_spill)(pset, thread);
+ }
+}
+
+processor_set_t
+task_choose_pset(
+ task_t task)
+{
+ processor_set_t pset = task->pset_hint;
+
+ if (pset != PROCESSOR_SET_NULL)
+ pset = choose_next_pset(pset);
+
+ return (pset);
+}
+
+/*
+ * Check for a preemption point in
+ * the current context.
+ *
+ * Called at splsched with thread locked.
+ */
+ast_t
+csw_check(
+ processor_t processor,
+ ast_t check_reason)
+{
+ processor_set_t pset = processor->processor_set;
+ ast_t result;
+
+ pset_lock(pset);
+
+ /* If we were sent a remote AST and interrupted a running processor, acknowledge it here with pset lock held */
+ bit_clear(pset->pending_AST_cpu_mask, processor->cpu_id);
+
+ result = csw_check_locked(processor, pset, check_reason);
+
+ pset_unlock(pset);
+
+ return result;
+}
+
+/*
+ * Check for preemption at splsched with
+ * pset and thread locked
+ */
+ast_t
+csw_check_locked(
+ processor_t processor,
+ processor_set_t pset,
+ ast_t check_reason)
+{
+ ast_t result;
+ thread_t thread = processor->active_thread;
+
+ if (processor->first_timeslice) {
+ if (rt_runq_count(pset) > 0)
+ return (check_reason | AST_PREEMPT | AST_URGENT);
+ }
+ else {
+ if (rt_runq_count(pset) > 0) {
+ if (BASEPRI_RTQUEUES > processor->current_pri)
+ return (check_reason | AST_PREEMPT | AST_URGENT);
+ else
+ return (check_reason | AST_PREEMPT);
+ }
+ }
+
+#if __SMP__
+ /*
+ * If the current thread is running on a processor that is no longer recommended,
+ * urgently preempt it, at which point thread_select() should
+ * try to idle the processor and re-dispatch the thread to a recommended processor.
+ */
+ if (!processor->is_recommended) {
+ return (check_reason | AST_PREEMPT | AST_URGENT);
+ }
+#endif
+
+ result = SCHED(processor_csw_check)(processor);
+ if (result != AST_NONE)
+ return (check_reason | result | (thread_eager_preemption(thread) ? AST_URGENT : AST_NONE));
+
+#if __SMP__
+ /*
+ * Same for avoid-processor
+ *
+ * TODO: Should these set AST_REBALANCE?
+ */
+ if (SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread)) {
+ return (check_reason | AST_PREEMPT);
+ }
+
+ /*
+ * Even though we could continue executing on this processor, a
+ * secondary SMT core should try to shed load to another primary core.
+ *
+ * TODO: Should this do the same check that thread_select does? i.e.
+ * if no bound threads target this processor, and idle primaries exist, preempt
+ * The case of RT threads existing is already taken care of above
+ */
+
+ if (processor->current_pri < BASEPRI_RTQUEUES &&
+ processor->processor_primary != processor)
+ return (check_reason | AST_PREEMPT);
+#endif
+
+ if (thread->state & TH_SUSP)
+ return (check_reason | AST_PREEMPT);
+
+#if CONFIG_SCHED_SFI
+ /*
+ * Current thread may not need to be preempted, but maybe needs
+ * an SFI wait?
+ */
+ result = sfi_thread_needs_ast(thread, NULL);
+ if (result != AST_NONE)
+ return (check_reason | result);
+#endif
+
+ return (AST_NONE);
+}
+
+/*
+ * set_sched_pri:
+ *
+ * Set the scheduled priority of the specified thread.
+ *
+ * This may cause the thread to change queues.
+ *
+ * Thread must be locked.
+ */
+void
+set_sched_pri(
+ thread_t thread,
+ int new_priority,
+ set_sched_pri_options_t options)
+{
+ thread_t cthread = current_thread();
+ boolean_t is_current_thread = (thread == cthread) ? TRUE : FALSE;
+ int curgency, nurgency;
+ uint64_t urgency_param1, urgency_param2;
+ boolean_t removed_from_runq = FALSE;
+
+ bool lazy_update = ((options & SETPRI_LAZY) == SETPRI_LAZY);
+
+ int old_priority = thread->sched_pri;
+
+ /* If we're already at this priority, no need to mess with the runqueue */
+ if (new_priority == old_priority)
+ return;
+
+ if (is_current_thread) {
+ assert(thread->runq == PROCESSOR_NULL);
+ curgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2);
+ } else {
+ removed_from_runq = thread_run_queue_remove(thread);
+ }
+
+ thread->sched_pri = new_priority;
+
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
+ (uintptr_t)thread_tid(thread),
+ thread->base_pri,
+ thread->sched_pri,
+ thread->sched_usage,
+ 0);
+
+ if (is_current_thread) {
+ nurgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2);
+ /*
+ * set_sched_pri doesn't alter RT params. We expect direct base priority/QoS
+ * class alterations from user space to occur relatively infrequently, hence
+ * those are lazily handled. QoS classes have distinct priority bands, and QoS
+ * inheritance is expected to involve priority changes.
+ */
+ uint64_t ctime = mach_approximate_time();
+ if (nurgency != curgency) {
+ thread_tell_urgency(nurgency, urgency_param1, urgency_param2, 0, thread);
+ }
+ machine_thread_going_on_core(thread, nurgency, 0, 0, ctime);
+ }
+
+ if (removed_from_runq)
+ thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ);
+ else if (thread->state & TH_RUN) {
+ processor_t processor = thread->last_processor;
+
+ if (is_current_thread) {
+ processor_state_update_from_thread(processor, thread);
+
+ /*
+ * When dropping in priority, check if the thread no longer belongs on core.
+ * If a thread raises its own priority, don't aggressively rebalance it.
+ * <rdar://problem/31699165>
+ */
+ if (!lazy_update && new_priority < old_priority) {
+ ast_t preempt;
+
+ if ((preempt = csw_check(processor, AST_NONE)) != AST_NONE)
+ ast_on(preempt);
+ }
+ } else if (!lazy_update && processor != PROCESSOR_NULL &&
+ processor != current_processor() && processor->active_thread == thread) {
+ cause_ast_check(processor);
+ }
+ }
+}
+
+/*
+ * thread_run_queue_remove_for_handoff
+ *
+ * Pull a thread or its (recursive) push target out of the runqueue
+ * so that it is ready for thread_run()
+ *
+ * Called at splsched
+ *
+ * Returns the thread that was pulled or THREAD_NULL if no thread could be pulled.
+ * This may be different than the thread that was passed in.
+ */
+thread_t
+thread_run_queue_remove_for_handoff(thread_t thread) {
+
+ thread_t pulled_thread = THREAD_NULL;
+
+ thread_lock(thread);
+
+ /*
+ * Check that the thread is not bound
+ * to a different processor, and that realtime
+ * is not involved.
+ *
+ * Next, pull it off its run queue. If it
+ * doesn't come, it's not eligible.
+ */
+
+ processor_t processor = current_processor();
+ if (processor->current_pri < BASEPRI_RTQUEUES && thread->sched_pri < BASEPRI_RTQUEUES &&
+ (thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor)) {
+
+ if (thread_run_queue_remove(thread))
+ pulled_thread = thread;
+ }
+
+ thread_unlock(thread);
+
+ return pulled_thread;
+}
+
+/*
+ * thread_run_queue_remove:
+ *
+ * Remove a thread from its current run queue and
+ * return TRUE if successful.
+ *
+ * Thread must be locked.
+ *
+ * If thread->runq is PROCESSOR_NULL, the thread will not re-enter the
+ * run queues because the caller locked the thread. Otherwise
+ * the thread is on a run queue, but could be chosen for dispatch
+ * and removed by another processor under a different lock, which
+ * will set thread->runq to PROCESSOR_NULL.
+ *
+ * Hence the thread select path must not rely on anything that could
+ * be changed under the thread lock after calling this function,
+ * most importantly thread->sched_pri.
+ */
+boolean_t
+thread_run_queue_remove(
+ thread_t thread)
+{
+ boolean_t removed = FALSE;
+ processor_t processor = thread->runq;
+
+ if ((thread->state & (TH_RUN|TH_WAIT)) == TH_WAIT) {
+ /* Thread isn't runnable */
+ assert(thread->runq == PROCESSOR_NULL);
+ return FALSE;
+ }
+
+ if (processor == PROCESSOR_NULL) {
+ /*
+ * The thread is either not on the runq,
+ * or is in the midst of being removed from the runq.
+ *
+ * runq is set to NULL under the pset lock, not the thread
+ * lock, so the thread may still be in the process of being dequeued
+ * from the runq. It will wait in invoke for the thread lock to be
+ * dropped.
+ */
+
+ return FALSE;
+ }
+
+ if (thread->sched_pri < BASEPRI_RTQUEUES) {
+ return SCHED(processor_queue_remove)(processor, thread);
+ }
+
+ processor_set_t pset = processor->processor_set;
+
+ rt_lock_lock(pset);
+
+ if (thread->runq != PROCESSOR_NULL) {
+ /*
+ * Thread is on the RT run queue and we have a lock on
+ * that run queue.
+ */
+
+ remqueue(&thread->runq_links);
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
+
+ thread->runq = PROCESSOR_NULL;
+
+ removed = TRUE;
+ }
+
+ rt_lock_unlock(pset);
+
+ return (removed);
projects / apple / xnu.git / blobdiff